ConeML

ConeML is a small-model research project focused on scratch training, curriculum geometry, and activation tuning.

The first public artifact is a 348M-parameter alpha model trained from scratch with a layered cone curriculum and staged SFT activation. The release documents what changed across stages, what activated in chat format, and what remains unresolved.

Current Focus

• Scratch-trained small language models
• Layered curriculum design
• Format activation and SFT staging
• Transparent internal probe reporting
• Reproducible model release artifacts

First Release

• Model: https://huggingface.co/ConeML/coneml-348m-alpha-polish900
• Website: https://coneml.com